Big Data Science in Finance

Chapter 2Neural Networks in Finance

Introduction

Neural networks are an important tool for machine learning. Truly deep learning was originally designed to model the complexities of the human brain. Neural networks typically require intensive computer power but with technology costs now at their historic low and projected to decrease further, neural networks are a cost-efficient yet powerful methodology for discovering nonlinear relationships that can be useful inputs into predicting future results. Here, following our paper, Aldridge and Avellaneda (2019), we discuss the theoretical background and develop a step-by-step implementation of a toy model for a neural network using financial data. The paper show practical and potentially profitable application of machine learning. The Appendix provides discussion and actual coding blocks for building a simple financial neural network in Python.

This chapter's focus is on simple explanation and the core principles of the neural network's design. One class of models that has been popular across image recognition and social media applications is Generative Adversarial Networks (GANs). The advantage of GANs is that they introduce randomization to enable classification of variables, even if none was previously available. Thus, Chen, Pelger, and Zhu (2019) use a deep learning GAN framework for estimating the stochastic discount factors (SDF), the unobservable Rosetta Stone of all pricing engines. As Chen, Pelger, and Zhu point out, SDF ...